Local AI line · stop 11 of 14 · 20 min · members
Four times the pixels without the plastic. Model choice, tiling and the sharpening you should not add.
Free with an account
Membership is free: an account opens all 86 script pages. The Lab, Studio Canvas and the paid guides need the $99 pass, paid once. Already signed in on this browser? The page opens by itself.
The problem
You wanted more pixels and you got a different image.
Upscaling models do not enlarge; they invent detail consistent with what they were trained on. That is why they work, and it is also why the result frequently looks over-sharpened, waxy, or subtly like a different photograph.
The failure is most visible in skin, which acquires a plastic smoothness, and in fine texture like fabric weave or grain, which gets replaced by a regular pattern.
Avoiding it is mostly about restraint: a smaller factor, the right model for the content, and a willingness to blend the result back towards the original.
Factor
The invention required grows with the factor.
At two times, the model is filling in modest gaps and the result generally stays faithful. At four, three quarters of the pixels are invented and the model's own habits dominate.
Where you genuinely need four, do it as two passes of two rather than one pass of four. The intermediate result constrains the second pass, and the outcome is usually more faithful than a single large jump.
Start from the largest native generation you can, so the upscale has less to do. Generating small and upscaling hard is the most common cause of a plastic result.
Model choice
A model trained on faces will apply faces to your fabric.
Upscaling models are specialised, and using one outside its territory produces characteristic artefacts. A general photographic model is the safe default; specialised ones do better on their subject and worse on everything else.
Test on a crop before committing to a full image. A hundred-pixel region tells you within seconds whether the model is inventing appropriately, and it costs nothing.
Judge at printed size or at the size the image will be used. Artefacts that dominate at full magnification are frequently invisible in the deliverable, and chasing them wastes time and often makes the image worse.
The blend
The single most effective technique, and the least used.
Enlarge the original conventionally to the target size, then place the upscaled version over it and reduce its opacity until the detail arrives without the texture.
Somewhere between half and three quarters is usually where it sits: enough added detail to justify the size, not enough of the model's invention to change the image.
Where only part of the frame needs it — a face that went plastic, a background that went busy — mask the blend rather than applying it uniformly. This is ordinary compositing and it solves most upscaling complaints outright.
Grain
Upscaling removes grain and replaces it with its own idea of texture.
If the image had film character, the upscale will have smoothed it and substituted something regular. That regularity is a large part of why upscaled images read as artificial.
Add grain as a final step, after the upscale and after the blend, at the size the image will be used. Grain applied before upscaling gets enlarged along with everything else and turns into blotches.
Match the grain to the size of the deliverable rather than to the pixel dimensions. Grain that looks correct at full magnification is invisible in print and vice versa.
When not to
It amplifies whatever is there, including the problems.
A frame that is out of focus, poorly composed or structurally wrong does not improve with more pixels. The model will confidently invent detail that was never captured, and the result is a large image that is still wrong.
Regenerate instead. It is faster than a long upscaling session and the outcome is honest.
The correct use is a good image that needs to be larger. That is the whole scope, and staying inside it is most of the skill.
1:1 · two hours